Search as Dialog

A search service today appears simple: a blank query box waiting for your search, and a resulting 10 blue links guiding into the World Wide Web. Yet its potential is unbounded. The empty box invites a dialog. The input is unrestricted, as if one were chatting with a very knowledgeable friend. The response can be anything -- just as if you were talking with that friend -- from a page with some links to tools and services that help you get something done. We like to think of it as the "Universal Knowledge Interface".

Today, however, search has it limits. It understands relatively little of what we say but is extremely useful for finding specific Web sites by name. According to our research, only 1 in 4 web queries you do gets you a satisfactory answer. Furthermore, a small fraction of the queries you do -- around 5 percent -- actually take up almost half the time you spend on searching! Indeed, today's search is good at yesterday's problem of finding things on the web, but isn't as good at helping you collect, organize, and act on information to make a decision.

Tomorrow, search will be the easiest way to answer a question or to complete a task. In the very near future search will have become such an essential companion that we will not understand how people survived without it -- indeed that trend is already happening.

Several accelerating trends will guide this trajectory. First, the Internet will continue to grow. Devices, users, information, and services are developing at double-digit rates with few signs of slowing. Second, the power of computer systems and algorithmic ingenuity brought to bear on navigating the online landscape will continue to defy the imagination. Third and finally, users will continue to demand ever more functionality from this boundless medium.

Crowd-Sourced Data and Services

One of the amazing phenomena of the last few years is the advent of large-scale, user-generated content sources, such as Wikipedia, Twitter, Flickr, and Facebook. These platforms combine the efforts of many individuals (a feat not possible at this scale prior to the Internet), but they are best experienced through a search interface that surfaces the highest-value content. Many start-ups companies now offer a pre-digested or filtered version of Twitter content that attempts to extract meaning from the stream.

We can expect this concept to transform an increasing number of information sources. Targeted blogging services are augmenting, and in some cases replacing, traditional, centralized news media outlets. Similarly, user contributed reviews are becoming an indispensable ingredient to a considered purchase of goods and services. The sheer volume of user contributed data is both a challenge and opportunity for engines: there is much to use but there is also a challenge in how best to use it to benefit the user.

Another form of user-generated data is the feedback user continually provide to Internet services through their behaviors -- sometimes explicitly, but often simply through their actions. For example, if almost every searcher chooses the second result for a query, that says a great deal about what the query actually means. Search services, such as Bing, currently use this data in aggregate to enhance search results and in many cases display 'search results' that look nothing like the 'ten blue links' to which we've become accustomed but rather in interfaces that best match what the user is trying to do. This is something we tried to do with Bing: a good example is when someone types in "flights from seattle to Chicago". We are pretty sure the "intent" of the user is to buy a ticket to get from Seattle to Chicago so we display some helpful information at the top of the page to tell the user if ticket prices are rising or falling. When you click on that, we take you to a customized interface that makes it easier to complete your task -- namely booking an airplane ticket.

By mining the vast amounts of behavioral data that accumulate through usage and through explicit and implicit contribution to the Web, search engines will become increasingly adept at anticipating user intentions. Ultimately this will extend to the common actions and services associated with the content someone is looking for. For example, it will be possible in the near future to reserve a table at a restaurant or order a taxi from the "search results" page for these queries. In the more distant future, your "virtual assistant" will be able to handle mundane tasks on your behalf and will become your broker to both the physical and virtual worlds because you've let it know so much about you.

Understanding Language and Context

Today, if you ask a search engine for "recent, positive reviews of the Amazon Kindle," it will completely ignore the nuances of the request because it understands relatively little of what we say. Search engines are very good at fuzzily matching the words of a query against their index of the web to create an appearance of understanding, much like the tourist who finds his way around Tokyo by matching ideograms on signs to cribbed directions. Increasingly, however, search engines will begin to understand more of the intention behind a user's query through the application of better web crawling and mining and natural-language-understanding algorithms. For example, search engines have historically successfully applied complex statistical analyses to the web in several languages to produce translators that handily beat traditional rule-based approaches.

We can expect these efforts to increase in sophistication, ultimately leading to engines that understand both the world and the structure of our language.

In addition to understanding language, search engines must correctly interpret context if they are to anticipate user needs. For example, the simple query, "traffic," has many possible intents. You could be looking for Internet traffic graphs, local road traffic maps, the cast of the movie Traffic, or songs from the band, Traffic. To resolve these possibilities, the engine must first understand that there at least four different 'things' in the real world that you could be talking about. Next, the engine might note the time of day (5pm, in this case), where I am (moving at 30 mph, so likely in a car and in Seattle near a freeway), and where I need to be (looking at my calendar and seeing that I have to be downtown by 6 p.m.). The combination of these factors will help the engine conclude that I need local traffic conditions rather than the band and it will be able to deliver useful knowledge to me instead of just information in the form of links on which I have to click.


Internet technology moves so rapidly that it's a real challenge to predict the future. However, it seems likely that the exponential growth patterns we are currently enjoying will continue for at least the next 10 to 20 years, which will fundamentally transform the information landscape. Search will be at the center of this transformation because it benefits from scale: It becomes better and more useful as the amount of data increases. It will also be enormously flexible by virtue of being able to understand language in ways that enable it to bridge the gap between physical and virtual worlds. In many ways, search technology is still in its infancy but much like a child, its potential is limitless.

We see Bing as the first step in this long process of transforming search from something which often points you somewhere else to try and find your answer. We see search as something that both understands you and the world in which it exists to provide you with insight and knowledge -- rather than more questions.